Workloads
Click a workload to expand its tasks| Workload | Current state | Opportunity to evaluate | Cost gain | ||||
|---|---|---|---|---|---|---|---|
| Model / stack | Quality | $/resolution | Model / stack | Quality | $/resolution | ||
Comparison setup and grade definitions
Baseline: GPT-4.1 for agent reasoning, extraction and summaries, with Claude Sonnet 4.6 for the optional in-product checker. Candidate: GPT-4.1 mini for selected extraction, confirmation and summary calls; GPT-4.1 retained for complex decisions; Claude Haiku 4.5 for the checker. These are examples to evaluate, not tested improvements.
Quality grade: A ≥95% pass; B ≥90%; C ≥80%; D <80%. Critical authorization, privacy or duplicate-action failures override the letter to Fail. The demo assumes none. The displayed percentage always belongs to the candidate.
Cost grade: A ≥30% reduction; B ≥15%; C ≥0%; D cost increase relative to baseline. A real report must also satisfy its quality comparison gate; eligibility is assumed here. Workload grades use cost per valid resolution. Task and call costs use attributed spend per invocation.
Rollup: total model spend across all attempts, including failures and retries, divided by valid resolutions. Quality is total valid resolutions divided by total attempts. Task and call replay sets are diagnostic examples and are not summed into full-workload costs.
Voice: the example speech stack is Deepgram Nova-3 for transcription and Cartesia Sonic 3 for synthesis. Audio quality, latency, speech charges, telephony and human costs are not evaluated here.